Professional Documents
Culture Documents
10.3934 Mbe.2017004
10.3934 Mbe.2017004
2017004
AND ENGINEERING
Volume 14, Number 1, February 2017 pp. 45–66
Luis L. Bonilla
G. Millán Institute, Fluid Dynamics, Nanoscience and Industrial Mathematics
Universidad Carlos III de Madrid
28911 Leganés, Spain
Vincenzo Capasso∗
ADAMSS, Universitá degli Studi di Milano
20133 MILANO, Italy
2010 Mathematics Subject Classification. Primary: 60G57, 60H10, 60H30, 60B10, 82C31,
92B05; Secondary: 60K35, 65C20, 65C35.
Key words and phrases. Angiogenesis, stochastic differential equations, birth and death pro-
cesses, growth processes, mean field approximation, hybrid models, propagation of chaos, ensemble
average.
∗ Corresponding author: Vincenzo Capasso.
45
46 BONILLA, CAPASSO, ALVARO, CARRETERO AND TERRAGNI
with the density of the full vessel network. In order to cope with anastomosis, the
vessel network has been modelled as a stochastic geometric process of Hausdorff
dimension one, as opposed to the system of tips which form a usual stochastic
particle system of Hausdorff dimension zero.
In [4] we have been able to derive (at least formally) a mean field equation for the
spatial density of tips, as a function of spatial location and velocity. This equation
is a parabolic integrodifferential equation of a Fokker-Planck type, having a source
term and a noninvertible diffusion matrix; it is second order in the derivatives with
respect to the velocities, and first order in the derivatives with respect to the position
coordinates. Apparently, together with the mean field equations for the underlying
fields, we have thus found an independent (deterministic) integrodifferential system
whose solution can provide the required (deterministic) kinetic parameters, which
drive the stochastic system for the tips, eventually leading to the stochastic vessel
network, at the microscale.
Mean field equations that follow from the “propagation of chaos” assumption
(law of large numbers) (see [29], [40], and references therein) are quite convenient
as they hold for any given realization of the underlying self-averaging stochastic
processes, but this requires a large number of tips at any location in the phase
space, and at any time.
Unfortunately anastomosis is responsible for a significant decay of the number
of tips, so to make inapplicable laws of large numbers on a single realization of
the stochastic process (a replica of the system), according to the “propagation of
chaos”.
However by re-examining the derivation given in [4], in [41] we have noticed
that the same deterministic description holds for vessel tip densities calculated by
averaging over replicas. This is close to J.W. Gibbs’ original ensemble average
interpretation of equilibrium statistical mechanics [23], except that, of course, our
system is always very far from equilibrium. In either cases, though with a different
interpretation of the solution, the deterministic description consists of a reaction-
diffusion equation for the tumor angiogenic factor (TAF) concentration coupled to
a Fokker-Planck type equation for the vessel tip density. To conclude, randomness
cannot be avoided; the deterministic description represents the evolution of an ideal
“mean” angiogenesis system and the evolution of a particular replica may deviate
significantly from it. These deviations are important and deserve further study, but
this is outside the scope of the present paper (see [8]).
The paper is organized as follows. Section 2 describes how our stochastic model
treats vessel branching, extension and anastomosis. In the subsequent three sections
we present the mathematical ingredients to be used in the sequel. In Section 6 we
derive the mean field approximation based on the propagation of chaos assumption.
In Section 7 a discussion is presented based on the outcomes of the numerical
simulations of the stochastic model as opposed to the mean field model. In Section 8
the ensemble average approach is presented to derive an equation of Fokker-Planck
type for the density of vessel tips and the TAF’s RDE. In Section 9 numerical
simulations of the deterministic averaged model are presented as opposed to the
stochastic model. Finally Section 10 contains our conclusions.
2. The mathematical model. Based on the above discussion, the main features
of the process of formation of a tumor-driven vessel network that we have considered
are (see [18, 30, 10])
48 BONILLA, CAPASSO, ALVARO, CARRETERO AND TERRAGNI
i) vessel branching;
ii) vessel extension;
iii) chemotaxis in response to a generic tumor angiogenic factor (TAF), released
by tumor cells;
iv) haptotactic migration in response to fibronectin gradient, emerging from the
extracellular matrix and through degradation and production by endothelial
cells themselves;
v) anastomosis, the coalescence of a capillary tip with an existing vessel.
Let N0 denote the initial number of tips, N (t) the number of tips at time t. Each
particle tip is characterized by its space Xi (t) and velocity vi (t) coordinates, so that
the whole process is characterized by the stochastic processes {(Xi (t), vi (t)), i =
1, ..., N (t), t ∈ R+ }. We model the capillary network of endothelial cells X(t) as the
union of all random trajectories representing the extension of individual capillary
tips from the (random) time of birth (branching) T i , to the (random) time of death
(anastomosis) Θi ,
N (t)
[
X(t) = {Xi (s), T i ≤ s ≤ min{t, Θi }}, (1)
i=1
giving rise to a stochastic network. It is clear that Xi (t) is a random closed set of
Hausdorff dimension zero, while X(t) is a random closed set of Hausdorff dimension
one.
We may describe both random sets by means of random distributions á la Dirac-
Schwarz [12]; δXi (t) (x) is the random distribution of the tip Xi (t) localized at point
x, for i = 1, . . . , N (t); δX(t) (x) is the random distribution of the whole network
X(t) localized at point x.
For convenience of the reader, we remind here that, under sufficient regularity
assumptions on a random set Ξ, it may admit a mean density λΞ , given by the
expectation of the random distribution [12] [1]
λΞ (x) = E[δΞ ](x). (2)
2.1. The underlying fields. TAF, fibronectin and matrix degrading enzymes ac-
tivate the migration of endothelial cells.
TAF diffuses, and it decreases where endothelial cells are present; we will adopt
the following model for the TAF evolution, according to which “consumption” (ac-
tually molecular binding) is due to the additional endothelial cells producing ves-
sels’ extension. Consumption is then taken proportional to the velocity vi , i =
1, . . . , N (t), of the tips, hence
∂
C(t, x) = d0 δA (x) + d2 4C(t, x)
∂t
N (t)
1 X i
− ηC(t, x) |v (t)|(KN ∗ δXi (t) )(x). (3)
N i=1
Parameters d2 , η > 0 denote the diffusivity and the rate of consumption, respec-
tively, while d0 represents the rate of production by a source located in a region
A ⊂ Rd , modelling e.g. a tumor mass. Later we will include the production term in
the evolution equation for C(t, x) only via the boundary conditions, meaning that
the tumor is located at the boundary of the relevant domain.
TUMOR-INDUCED ANGIOGENESIS 49
All these (random) partial differential equations are subject to suitable boundary
and initial conditions.
For the sake of technical simplification, in the sequel we will not consider the last
two underlying fields f and m.
4. The dynamics.
4.1. Tip branching. Two kinds of branching have been identified; either from a
tip or from a mature vessel; here for the sake of simplicity, we shall consider only
tip branching.
The birth process of new tips can be described in terms of a marked point process
(see e.g. [5]), by means of the random measure Φ on BR+ ×Rd ×Rd such that, for any
t ≥ 0 and any B ∈ BRd ×Rd ,
Z tZ
Φ((0, t] × B) := Φ(ds × dx × dv), (8)
0 B
where Φ(ds × dx × dv) is the random variable that counts those tips born from an
existing tip during times in (s, s + ds], with positions in (x, x + dx], and velocities
in (v, v + dv].
We assume that the probability that a tip branches from one of the N (t) ex-
isting ones during an infinitesimal time interval (t, t + dt] is function of the TAF
concentration C(t, x)
N (t)
X
prob(N (t + dt) − N (t) = 1|Ft− ) = α(C(t, Xi (t))) dt, (9)
i=1
Z tZ
Ψ((0, t] × B) := Ψ(ds × dx × dv) (14)
0 B
where Ψ(ds × dx × dv) is the random variable counting those tips that are absorbed
by the existing vessel network during time (s, s + ds], with position in (x, x + dx],
and velocity in (v, v + dv].
Thanks to the choice of a Langevin model for the vessels extension, we may
assume that the trajectories are sufficiently regular and have an integer Hausdorff
dimension 1.
Hence [12] the random measure
A ∈ BRd 7→ H1 (X(t) ∩ A) ∈ R+ (15)
admits a random generalized density δX(t) (x) with respect to the usual Lebesgue
measure on Rd such that, for any A ∈ BRd ,
Z
H1 (X(t) ∩ A) = δX(t) (x)dx. (16)
A
Given the definition of X(t) as the union of all trajectories up to time t, the
Hausdorff measure associated with the stochastic network X(t) can be expressed in
terms of the occupation time of a region A ∈ BRd by tips existing up to time t > 0
(see [32], [26])
Z t N (s)
X
H1 (X(t) ∩ A) = γ ds IA (Xi (s))
0 i=1
Z t N (s)
X
= γ ds Xi (s) (A), (17)
0 i=1
We assume that the probability per unit time that a typical tip Xk (t) meets the
existing vessel network X(t) (and “dies”) is a linear function of the scaled random
distribution N −1 δX(t) of the stochastic network at point Xk (t) and time t, so that
the probability that an active tip dies during an infinitesimal time interval is given
by
N (t)
X
prob(N (t + dt) − N (t) = −1|Ft− ) = N −1 (KN ∗ δX(t) )(Xi (t))dt. (19)
i=1
We then claim that the compensator of the random measure Ψ(ds × dx × dv) is
N −1 (KN ∗ δX(s) )(x)N QN (s)(d(x, v))ds. (20)
52 BONILLA, CAPASSO, ALVARO, CARRETERO AND TERRAGNI
6. Mean field approximation. By Itô’s formula (see e.g. [9], p. 270), we may
obtain the evolution equation of the random measure (QN (t))t∈R+ as follows. For
any test function g ∈ Cb (Rd × Rd ),
Z Z
g(x, v) QN (t)(d(x, v)) = g(x, v) QN (0)(d(x, v))
Z tZ
+ v · ∇x g(x, v)QN (s)(d(x, v)) ds
0
Z tZ
+ [F(C(s, x)) − kv]·∇v g(x, v)QN (s)(d(x, v)) ds
0
Z tZ 2
σ
+ ∆v g(x, v)QN (s)(d(x, v))ds
0 2
Z tZ
+ α(C(s, x))δv0 (v)g(x, v)QN (s)(d(x, v)) ds
0
Z tZ
1
− (KN ∗ δX(s) )(x)g(x, v)QN (s)(d(x, v)) ds
0 N
+M̃N (t), (21)
where
M̃N (t) = M̃1,N (t) + M̃2,N (t) + M̃3,N (t),
with
t N (s)
1 X
Z
M̃1,N (t) = ∇v g(Xk (s), vk (s)) · dWk (s), (22)
0 N
k=1
Z tZ
M̃2,N (t) = g(x, v)[ΦN (ds × dx × dv)
0
−α(C(s, x))δv0 (v)QN (s)(d(x, v))ds], (23)
TUMOR-INDUCED ANGIOGENESIS 53
Z tZ
M̃3,N (t) = g(x, v)[ΨN (ds × dx × dv)
0
1
− (KN ∗ δX(s) )(x)QN (s)(d(x, v))ds . (24)
N
All (M̃1,N (t))t∈R+ , (M̃2,N (t))t∈R+ , and (M̃3,N (t))t∈R+ are zero mean martingales,
so that (M̃N (t))t∈R+ is itself a zero mean martingale.
By Doob’s inequality
P
M̃N (t) −→ 0.
N →∞
This is the substantial reason of the possible deterministic limiting behavior of
the process, as N → ∞.
As a consequence, if we suppose that indeed, for a large value of the scale pa-
rameter N, we may claim
QN (t)(d(x, v)) → Q∞ (t)(d(x, v)) = p(t, x, v)dx dv. (25)
1
Assuming that (25) holds true, the scaled stochastic distribution of vessels (KN ∗
N
δX(t) )(x) can be itself approximated by the mean value
Z t
λ(t, x) = E[δX(t) ](x) = γ p̃(s, x) ds, (26)
0
where Z
p̃(t, x) = E[δXi (t) ](x) = p(t, x, v0 ) dv0 (27)
so that the evolution equation (3) may be approximated by its (deterministic) mean
field version
∂ e
C(t, x) = d2 ∆x C(t,
e x) − η C(t,
e x)|j(t, x)|. (29)
∂t
TAF injection from the tumor is realized as a nonzero flux boundary condition for
this equation [4], instead of including it explicitly as in (3).
Based on the above discussion, one might claim that the evolution equation for
the density p(t, x, v) is the following one
∂ α1 C(t,
e x)
p(t, x, v) = p(t, x, v)δv0 (v)
∂t CR + C(t,
e x)
Z t
−γp(t, x, v) p̃(s, x) ds
0
−v · ∇x p(t, x, v)
+k∇v · (vp(t, x, v))
" #
∇x C(t,
e x)
−d1 ∇v · p(t, x, v)
e x)]q
[1 + γ1 C(t,
σ2
+ ∆v p(t, x, v). (30)
2
The coupled system (29), (30) is subject to suitable boundary and initial condi-
tions, to be discussed later.
This approach is called “hybrid”, since we have substituted all stochastic un-
derlying fields by their “mean” counterparts; most of the current literature could
now be reinterpreted along these lines. Indeed, one should check that the hybrid
system is fully compatible with a rigorous derivation of the evolution for the ves-
sel densities. Nonlinearities in the full model are a big difficulty in this direction;
a rigorous derivation of the convergence results requires a nontrivial mathematical
analysis, which is under investigation; for this it is instrumental a proof of existence,
uniqueness and sufficient regularity of the solution of the mean field equations (see
[29], and [6]). We wish to stress that anyhow substituting mean geometric densities
of tips or of full vessels to the corresponding stochastic quantities leads to an ac-
ceptable coefficient of variation (percentage error) only when a law of large numbers
can be applied, i.e. whenever the relevant numbers per unit volume are sufficiently
large; otherwise stochasticity cannot be avoided, and in addition to mean values,
the mathematical analysis and/or simulations should provide confidence bands for
all quantities of interest (see e. g. [7]). An interesting case in this direction is
discussed in [11].
tip density based on self-averaging due to the law of large numbers does not seem
correct.
1 1 1.2
1
0.8 0.8
0.8
0.6 0.6
0.6
0.4 0.4
0.4
0.2 0.2
0.2
0 0 0
−0.2 −0.2 −0.2
1 1 1
0.5 1 0.5 1 0.5 1
0.8 0.8 0.8
0 0 0
y/L −0.5 0.4
0.6
y/L −0.5 0.4
0.6
y/L −0.5 0.4
0.6
−1 0
0.2
x/L −1 0
0.2
x/L −1 0
0.2
x/L
Here the number of tips at time t, N (t, ω), may be different for different replicas.
Similarly at time t, the stochastic distribution of tips per unit volume in physical
space is given by
N (t,ω)
X
Q̃∗N (t, x, ω) = δXi (t,ω) (x), (32)
i=1
56 BONILLA, CAPASSO, ALVARO, CARRETERO AND TERRAGNI
and
Z t NX
(s,ω)
δX(t,ω) (x) := γ δXi (s,ω) (x)ds (33)
0 i=1
represents the concentration of all vessels per unit volume in physical space, at time
t, i.e., the vessel network.
By following a similar approach as in our previous paper [4], we may then ob-
tain the weak formulation of the stochastic evolution of Q∗N (t, x, v, ω), by following
similar steps as in Equation (21):
Z Z
∗
g(x, v) QN (t, x, v)dxdv = g(x, v) Q∗N (0, x, v)dxdv
Z tZ
+ v · ∇x g(x, v)Q∗N (s, x, v)dxdv ds
0
Z tZ
+ [F(C(s, x)) − kv]·∇v g(x, v)Q∗N (s, x, v)dxdv ds
0
Z tZ 2
σ
+ ∆v g(x, v)Q∗N (s, x, v)dxdvds
0 2
Z tZ
+ α(C(s, x))δv0 (v)g(x, v)Q∗N (s, x, v)dxdv ds
0
Z tZ
− δX(s) (x)g(x, v)Q∗N (s, x, v)dxdv ds
0
+ M N (t), (34)
where we have omitted the variable ω not to encumber the formulas. Here M N (t) is
again a zero mean martingale that collects all source of randomness of the system,
g(x, v) is a smooth test function.
By mimicking the typical kernel density estimation approach in Statistics (see
e.g. [33], page 489), we introduce the (random) empirical distribution of tips per
unit volume in the (x, v) phase space,
N
1 X
pN(t, x, v) = (KN ∗ Q∗N (t, ·, ·, ω))(x, v). (35)
N ω=1
Here the mollifying kernel KN (x, v) tends to a Dirac delta function δ0,0 as N → ∞.
Correspondingly, the empirical distribution of tips per unit volume in the physical
space and the vessel tip flux are,
N
1 X e ∗N (t, ·, ω))(x),
p̃N (t, x) = (KN ∗ Q (36)
N ω=1
N N (t,ω)
1 X X i
jN (t, x) = v (t, ω)(KN ∗ δXi (t,ω) )(x), (37)
N ω=1 i=1
respectively. We now conjecture that the following limit exists
p(t, x, v) = lim pN (t, x, v), (38)
N →∞
and that p(t, x, v) is the deterministic distribution of tips per unit phase space
volume. The proof of this statement is not trivial (and out of the scope of this
paper), as the usual assumptions on the kernel density estimation (see page 489 of
TUMOR-INDUCED ANGIOGENESIS 57
[33]) may not apply. Correspondingly the deterministic distribution of tips per unit
volume in the physical space should exist as the following limit
Finally, we may obtain the deterministic version of the vessel tip flux as
Figure 2 shows the marginal tip density p̃(t, x, y) ≈ p̃N (t, x, y) calculated from
(36) with N = 50 replicas at the times represented in Figure 1. The tips proliferate
after a few hours and reach a high value by branching onto the free space ahead
of them. Influenced by chemotaxis, the marginal tip density thickens about the x
axis and it forms a lump that advances toward the tumor. Behind the lump, the
density drops to a low value as few active tips remain. As shown by Figure 1, the
network of vessels is quite dense in the wake of the tips and therefore anastomosis
reduces enormously the number of active tips there. They are numerous only at
the leading part of the lump where free space is available. Thus the definition
of marginal tip density based on ensemble average over replicas provides a better
deterministic description of angiogenesis than the density based on the law of large
numbers. There are no tips or very few ones in large regions of the physical space.
58 BONILLA, CAPASSO, ALVARO, CARRETERO AND TERRAGNI
we may expect that the ensemble average of the stochastic equation (34) tends in
its strong form to the same equation of Fokker-Planck type as (30):
∂ α1 C(t, x)
p(t, x, v) = p(t, x, v)δv0 (v)
∂t CR + C(t, x)
Z t
− γp(t, x, v) p̃(s, x) ds − v · ∇x p(t, x, v)
0
∇x C(t, x)
+ k∇v · [vp(t, x, v)] − d1 ∇v · p(t, x, v)
[1 + γ1 C(t, x)]q
σ2
+ ∆v p(t, x, v). (41)
2
Here the marginal vessel tip density,
Z
p̃(t, x) = p(t, x, v0 ) dv0 , (42)
A crucial point in the derivation of (41) as the ensemble average of Equation (34)
is the approximation of nonlinear terms in the latter, e.g. the anastomosis term.
Had the law of large numbers been applicable (on the single replica, according to
the propagation of chaos assumption), the mean values of nonlinear terms could be
factorized. With the ensemble average description, factorization of nonlinear terms
requires further investigation, and this is out of the scope of the present paper. By
assuming factorization occurs, we may expect that the ensemble average of Equation
(34) tends in its strong form to (41).
By an abuse of notations, we are using in Equation (43) the same letter C for
the deterministic counterpart of the TAF field, too.
We need to stress here that the p(t, x, v) appearing in Equation (30) represents
a scaled density in phase space, the integral of which over the whole phase space
is N (t)/N , where N (t) is the mean number of tips at time t. This number coin-
cides with the number of tips in a single realization of our supposed self-averaging
stochastic process. In contrast to this, the p(t, x, v) appearing in Equation (41)
represents the true concentration per unit volume of phase space.
For appropriate initial and boundary data, it is possible to prove that (41) and
(43) have a unique smooth solution [16].
TUMOR-INDUCED ANGIOGENESIS 59
8.1. Boundary and initial conditions. We have solved the system of equations
(41) and (43) in a two dimensional strip geometry using the initial and boundary
conditions introduced in [4]. The strip is Ω = [0, L] × R ⊂ R2 , its left boundary
Ω0 = (0, y), y ∈ R, is the primary vessel issuing new vessels, and ΩL = (L, y),
y ∈ R, represents the tumor which is a source of the TAF C. Let c1 (y) be the TAF
flux emitted by the tumor at x = L. As boundary conditions for the TAF we have
taken
∂ ∂ c1 (y)
C(t, 0, y) = 0, C(t, L, y) = , (45)
∂x ∂x d2
and C → 0 as |y| → ∞. We have used a Gaussian as the initial condition for the
TAF
2
/c2 +y 2 /b2 ]
C(0, x, y) = 1.1 CR e−[(x−L) , (46)
for appropriate b and c.
As boundary conditions for the tip density we have taken
k|v−v0 |2
e− σ2
p+ (t, 0, y, v, w) = R R k|v0 −v0 |2
∞ ∞
0 −∞
v 0 e− σ2dv 0 dw0
Z 0 Z ∞
0 − 0 0 0 0
× j0 (t, y)− v p (t, 0, y, v , w )dv dw , (47)
−∞ −∞
k|v−v0 |2
− e− σ2
p (t, L, y, v, w) = R R k|v0 −v0 |2
0 ∞
−∞ −∞
e− σ2dv 0 dw0
Z ∞Z ∞
+ 0 0 0 0
× p̃(t, L, y)− p (t, L, y, v , w )dv dw , (48)
0 −∞
p(t, x, v) → 0 as |v| → ∞, (49)
+ −
where p = p for v > 0 and p = p for v < 0, v = (v, w). These phenomenological
conditions give the tip density for velocities entering the slab region in terms of the
tip density for velocities exiting the slab region. They are constructed so that they
give the marginal tip density at x = L and the tip flux density at x = 0, which is
[4]
v0 L
j0 (t, y) = p 2 α(C(t, 0, y)) p(t, 0, y, v0 , w0 ), (50)
v0 + w02
for the vector velocity v0 = (v0 , w0 ). Improving the boundary conditions requires a
separate and detailed work comparing stochastic and deterministic descriptions of
the angiogenesis process.
Finally as initial condition for the tip density we have taken
2 2 N
2e−x /lx −|v−v0 |2 /σv2 X 0
1 −|y−yi |2 /ly2
p(0, x, y, v, w) = e √ e . (51)
π 3/2 lx σv2 i=1
πly
where δvσ0v (v) is the mollified version of the usual Dirac delta by a Gaussian kernel.
This initial condition corresponds to the following initial condition for the stochastic
60 BONILLA, CAPASSO, ALVARO, CARRETERO AND TERRAGNI
process: There are N0 equally spaced initial tips at x = 0, with vertical positions
yi equally spaced on the interval [−Ly , Ly ], whose initial velocities are normally
distributed about v0 with standard deviation σv .
δ β A Γ1 κ Γ χ
d1 CR kL α1 L d2 γ η
ṽ02 ṽ0 γ1 CR
ṽ03 ṽ0 Lṽ02L
1.5 5.88 22.42 0.145 1 0.0045 0.002
Table 1. Dimensionless parameters. ṽ0 = 40µm/h is a typical cell velocity.
800 400
600 300
400 200
200 100
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x/L x/L
400
300
300
200
200
100
100
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x/L x/L
100 200
80
150
60
100
40
20 50
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x/L x/L
200
150
150
100
100
50
50
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x/L x/L
10. Conclusions. In our recent papers [4] and [41] we have explored the behavior
of a stochastic angiogenesis model, and of its possible deterministic approximation.
In this model, the tips undergo a stochastic process of tip branching, vessel extension
and anastomosis whereas TAF is described by a reaction-diffusion equation with a
sink term proportional to the tip flow.
In [4] the empirical measure describing the tip distribution had been assumed
to satisfy a “law of large numbers” for any single replica of the process, i. e. the
classical “propagation of chaos” assumption (see e.g. [40], and references therein),
so that it admits a position-velocity density which is shown to satisfy a nonlin-
ear integro-differential equation of a Fokker-Planck type, coupled with a reaction-
diffusion equation for the TAF concentration, in which the stochastic tip flow has
been replaced by its mean field approximation, deriving from the tip mean density.
On the other hand, in [41] we have solved numerically the stochastic model for
many realizations (independent replicas of the system). Numerically calculated
velocity fluctuations have revealed that they do not decay even as the number
of initial vessel tips increases. This shows that the stochastic model is not self-
averaging and therefore we cannot use the “propagation of chaos” assumption to
derive a mean field deterministic approximation of the stochastic model. The main
reason being that anastomosis eliminates many vessel tips, resulting in the fact that
there never are enough tips for a law of large numbers apply on a single replica. The
vessel network has shown to be quite different for different replicas of the stochastic
process.
However by re-examining the derivation given in [4], we conclude that the same
deterministic description holds for vessel tip densities calculated by averaging over
replicas. The deterministic description consists of a reaction-diffusion equation for
the TAF concentration coupled to a Fokker-Planck type equation for the vessel tip
density. The latter contains a birth term corresponding to tip branching and a
death integral term corresponding to anastomosis or tip merging. The coefficient
of the latter term has been fitted by comparison with the stochastic description:
optimal selection produces a good fit for the evolution of the total number of tips.
For the averaged deterministic reaction-diffusion system, boundary conditions
have been proposed in [41], which describe the flux of vessel tips injected from a
primary blood vessel in response to TAF emitted by the tumor and the tip density
eventually arriving at the tumor. Numerical solution of the model in a simple
geometry shows how tips are created at the primary blood vessel, propagate and
proliferate towards the tumor and may or not reach it after a certain time depending
on the parameter values. This is consistent with known biological facts.
Actually nontrivial additional investigation is required for a rigorous derivation
of the deterministic approximation of the relevant empirical measures and their evo-
lution equations. Anyhow we wish to convey a general message elicited by the pro-
posed angiogenesis model: in stochastic models containing birth and death processes
in addition to Brownian motion (Langevin equations), the death processes may pre-
clude reaching the large number of individuals required to have self-averaging and
64 BONILLA, CAPASSO, ALVARO, CARRETERO AND TERRAGNI
REFERENCES
[1] L. Ambrosio, V. Capasso and E. Villa, On the approximation of mean densities of random
closed sets, Bernoulli, 15 (2009), 1222–1242.
[2] A. R. A. Anderson and M. A. J. Chaplain, Continuous and discrete mathematical models of
tumour-induced angiogenesis, Bull. Math. Biol., 60 (1998), 857–900.
[3] C. Birdwell, A. Brasier and L. Taylor, Two-dimensional peptide mapping of fibronectin from
bovine aortic endothelial cells and bovine plasma, Biochem. Biophys. Res. Commun., 97
(1980), 574–581.
[4] L. L. Bonilla, V. Capasso, M. Alvaro and M. Carretero, Hybrid modelling of tumor-induced
angiogenesis, Physical Review E, 90 (2014), 062716.
[5] P. Bremaud, Point Processes and Queues. Martingale Dynamics, Springer-Verlag, New-York,
1981.
[6] M. Burger, V. Capasso and D. Morale, On an aggregation model with long and short range
interactions, Nonlinear Anal. Real World Appl., 8 (2007), 939–958.
[7] M. Burger, V. Capasso and L. Pizzocchero, Mesoscale averaging of nucleation and growth
models, Multiscale Modeling and Simulation, 5 (2006), 564–592.
[8] V. Capasso, Randomness and Geometric Structures in Biology, in Pattern Formation in
Morphogenesis. Problems and Mathematical Issues (eds. V. Capasso, M. Gromov, A. Harel-
Bellan, N. Morozova and L.L. Pritchard), Springer, Heidelberg, 2013.
[9] V. Capasso and D. Bakstein, An Introduction to Continuous-time Stochastic Processes, 3rd
edition, Birkhäuser, Boston, 2015.
[10] V. Capasso and D. Morale, Stochastic modelling of tumour-induced angiogenesis, J. Math.
Biol., 58 (2009), 219–233.
TUMOR-INDUCED ANGIOGENESIS 65
[11] V. Capasso, D. Morale and G. Facchetti, The Role of Stochasticity for a Model of Retinal
Angiogenesis, IMA J. Appl. Math., 77 (2012), 729–747.
[12] V. Capasso and E. Villa, On the geometric densities of random closed sets, Stoch. Anal. Appl.,
26 (2008), 784–808.
[13] P. F. Carmeliet, Angiogenesis in life, disease and medicine, Nature, 438 (2005), 932–936.
[14] P. Carmeliet and R. K. Jain, Molecular mechanisms and clinical applications of angiogenesis,
Nature, 473 (2011), 298–307.
[15] P. Carmeliet and M. Tessier-Lavigne, Common mechanisms of nerve and blood vessel wiring,
Nature, 436 (2005), 193–200.
[16] A. Carpio and G. Duro, Well posedness of a kinetic model for angiogenesis, Nonlinear Anal-
ysis; Real World Applications, 30 (2016), 184–212.
[17] N. Champagnat and S. Méléard, Invasion and adaptive evolution for individual-based spatially
structured populations, J. Math. Biol., textbf55 (2007), 147–188.
[18] M. Chaplain and A. Stuart, A model mechanism for the chemotactic response of endothelial
cells to tumour angiogenesis factor, IMA J. Math. Appl. Med. Biol., 10 (1993), 149–168.
[19] M. Corada, L. Zanetta, F. Orsenigo, F. Breviario, M. G. Lampugnani, S. Bernasconi, F.
Liao, D. J. Hicklin, P. Bohlen and E. Dejana, A monoclonal antibody to vascular endothelial-
cadherin inhibits tumor angiogenesis without side effects on endothelial permeability. Blood,
100 (2002), 905–911.
[20] S. L. Cotter, V. Klika, L. Kimpton, S. Collins and A. E. P. Heazell, A stochastic model
for early placental development, J. R. Soc. Interface, 11 (2014), 20140149, Available from:
http://rsif.royalsocietypublishing.org/content/11/97/20140149.
[21] R. F. Gariano and T. W. Gardner, Retinal angiogenesis in development and disease, Nature,
438 (2004), 960–966.
[22] J. Folkman, Tumour angiogenesis, Adv. Cancer Res., 19 (1974), 331–358.
[23] J. W. Gibbs, Elementary Principles of Statistical Mechanics, Yale Bicentennial Publications,
Scribner and Sons, New York, 1902.
[24] H. A. Harrington, M. Maier, L. Naidoo, N. Whitaker and P. G. Kevrekidis, A hybrid model
for tumor-induced angiogenesis in the cornea in the presence of inhibitors, Mathematical and
Computer Modelling, 46 (2007), 513–524.
[25] R. K. Jain and P. F. Carmeliet, Vessels of death or life, Sci. Am., 285 (2001), 38–45.
[26] S. Karlin and H. M. Taylor, A Second Course in Stochastic Processes, Academic Press, New
York, 1981.
[27] N. V. Mantzaris, S. Webb and H. G. Othmer, Mathematical modeling of tumor-induced
angiogenesis, J. Math. Biol., 49 (2004), 111–187.
[28] D. Morale, V.Capasso and K.Ölschlaeger, An interacting particle system modelling aggrega-
tion behavior: From individuals to populations, J. Math. Biol., 50 (2005), 49–66.
[29] K. Oelschläger, On the derivation of reaction-diffusion equations as limit dynamics of systems
of moderately interacting stochastic processes, Probab. Theor. Relat. Fields, 82 (1989), 565–
586.
[30] M. J. Plank and B. D. Sleeman, Lattice and non-lattice models of tumour angiogenesis, Bull.
Math. Biol., 66 (2004), 1785–1819.
[31] M. Hubbard, P. F. Jones and B. D. Sleeman, The foundations of a unified approach to
mathematical modelling of angiogenesis, Int. J. Adv. Eng. Sci. and Appl. Math., 1 (2009),
43–52.
[32] P. E. Protter, Stochastic Integration and Differential Equations, Second Edition, Springer-
Verlag, Heidelberg, 2004.
[33] G. G. Roussas, A Course in Mathematical Statistics, 2nd edition, Academic Press, San Diego,
CA, 1997.
[34] M. Scianna, L. Munaron and L. Preziosi, A multiscale hybrid approach for vasculogenesis and
related potential blocking therapies, Prog. Biophys. Mol. Biol., 106 (2011), 450–462.
[35] A. Stéphanou, S. R. McDougall, A. R. A. Anderson and M. A. J. Chaplain, Mathematical
modelling of the influence of blood rheological properties upon adaptative tumour-induced
angiogenesis, Math. Comput. Modelling, 44 (2006), 96–123.
[36] C. L. Stokes and D. A. Lauffenburger, Analysis of the roles of microvessel endothelial cell
random motility and chemotaxis in angiogenesis, J. Theor. Biol., 152 (1991), 377–403.
[37] S. Sun, M. F. Wheeler, M. Obeyesekere and C. W. Patrick Jr., A deterministic model of
growth factor-induced angiogenesis, Bull. Math. Biol., 67 (2005), 313–337.
66 BONILLA, CAPASSO, ALVARO, CARRETERO AND TERRAGNI